MCP server for AutoMem: AI memory storage and recall
2026-11-16
星數
—
工具
—
可用率
—
The Problem We Solve
Every AI conversation starts from zero. Claude forgets your coding style. Cursor can't learn your patterns. Your assistant doesn't remember yesterday's decisions.
Until now.
AutoMem MCP connects your AI to persistent memory powered by AutoMem - a graph-vector memory service.
What You Get
🧠 Persistent Memory Across Sessions
AI remembers decisions, patterns, and context forever
Works across all MCP platforms - Claude Desktop, Cursor, Claude Code, OpenAI Codex, OpenClaw, Hermes, Grok Build, Google Antigravity
Cross-device sync - same memory on Mac, Windows, Linux
🏆 Graph-Vector Architecture
11 public authorable relationship types between memories (recall results may also include read-only system/internal relations that are not valid associate_memories inputs)
Claude Desktop will prompt you for your AutoMem API URL (http://127.0.0.1:8001 for local)
Optionally enter your API Key (required for Railway, skip for local)
Click Enable
The bundle includes a fixed version of the MCP server. To update, download and install the latest .mcpb again; restarting Claude or updating the npm package does not upgrade an installed bundle. See updating Claude Desktop for version checks and migration from JSON configuration. The AutoMem backend runs separately.
Then add the paste-ready Personal Preferences starter from templates/CLAUDE_DESKTOP_INSTRUCTIONS.md. That's it: Claude now has persistent memory and knows when to use it.
Other Platforms
Connect your AI tools to the AutoMem service you just started.
# Guided install - pick where AutoMem runs, verify it, write .env, and
# configure your agents (Codex, Claude Code, Cursor, OpenClaw, Hermes)
npx @verygoodplugins/mcp-automem install
Every change is shown in a review plan before anything is written, and each
modified file keeps a .bak backup. Add --dry-run to preview, --yes to
apply non-interactively. See the Installation Guide
for all flags.
setup is an alias of install. To print config snippets again later, use npx @verygoodplugins/mcp-automem config.
3. Platform-Specific Setup
For Claude Code (plugin — recommended):
# In Claude Code:
/plugin marketplace add verygoodplugins/mcp-automem
/plugin install automem@verygoodplugins-mcp-automem
Claude Code prompts for your AutoMem URL and API key at enable time, bundles the MCP server and silent recall/store-tracking hooks, and auto-updates. Prefer hooks and permissions written directly into ~/.claude/ instead? Run npx @verygoodplugins/mcp-automem claude-code.
On Windows, the hook payload assumes a POSIX shell environment such as Git Bash, MSYS2, or WSL — only bash is required (the hooks are pure bash+sed).
For Cursor IDE:
Run the guided installer from your project and select Cursor:
npx @verygoodplugins/mcp-automem install
This connects Cursor to your AutoMem service, installs project memory rules, and registers a session-start recall hook. Already have an endpoint? Use npx @verygoodplugins/mcp-automem cursor --endpoint https://your-automem.example.
The one-click button configures only the MCP server for a local AutoMem instance. Update its endpoint/key in Cursor if needed, then run the CLI above to add project rules and the hook:
An optional sidecar service (deployable to Railway or any Docker host) connects AutoMem to platforms that support remote MCP over Streamable HTTP or SSE — ChatGPT (Developer Mode connectors), Claude.ai web and Claude Mobile, and ElevenLabs Agents.
👉 Remote MCP setup for deployment, connect URLs, and per-platform screenshots.
store_memory — Save memories with content, tags, importance, metadata. Two modes:
Single (default): top-level content plus optional fields, including embedding, t_valid, t_invalid, custom id.
Batch: memories: [...] (≤500 items) for bulk ingestion. Per-item id/embedding/t_valid/t_invalid are not supported in batch mode.
recall_memory — Three modes selected by which params you pass:
ID fetch: memory_id → fetches one memory by ID; updates last_accessed.
Tag enumeration: tags + exhaustive: true → paginated exact-match listing for cleanup/audit workflows where ranked recall undercounts. Pair with limit (≤200) and offset; returns has_more, plus next_offset for the next page while more remain.
Ranked retrieval (default): hybrid search across vector, keyword, tags, recency/state controls, score filters, and graph expansion. Supports state_mode, recency_bias, scope_fallback, expand_respect_tags, min_score, adaptive_floor, and diagnostics such as tag_scope, score_filter, query_time_ms, vector_search, and per-result outside_tag_scope/state_replaces.
associate_memories — Create relationships (11 public authorable types; recall results may also include read-only system relations). Supports single-pair mode and batch mode via associations: [...] (≤500) with relation-specific props like reason, context, resolution, observations, transformation, and role.
update_memory — Modify existing memories
delete_memory — Two modes:
Single (default): memory_id → removes one memory and its embedding.
Bulk-by-tag: tags: [...] → bulk-delete all memories matching ANY tag (exact, case-insensitive). No dry-run; verify with recall_memory({ tags, exhaustive: true }) first.
check_database_health — Monitor service health, degraded state, sync counts, vector dimensions, and enrichment diagnostics when the service provides them
Advanced Recall (v0.8.0+)
Multi-hop Reasoning - Answer complex questions like "What is Amanda's sister's career?"
mcp__memory__recall_memory({
query: "What is Amanda's sister's career?",
expand_entities: true, // Finds "Amanda's sister is Rachel" → memories about Rachel
});
Context-Aware Coding - Recall prioritizes language and style preferences
A-MEM (July 2025): Dynamic memory organization with Zettelkasten principles
MELODI (DeepMind, 2025): 8x memory compression without quality loss
ReadAgent (DeepMind, 2024): 20x context extension through gist memories
This MCP package provides the bridge between your AI and that research-validated memory system. The backend has also been benchmarked on the neutral Agent Memory Benchmark, including BEAM large-context scaling tiers — reproducible end to end, so you can run it yourself.
Community & Support
💬 Discord - Join the community, get help, share feedback